Suppose vector \theta
is all the parameters in a neural network, I wonder how to compute hessian matrix for \theta
in pytorch.
Suppose the network is as follows:
class Net(Module):
def __init__(self, h, w):
super(Net, self).__init__()
self.c1 = torch.nn.Conv2d(1, 32, 3, 1, 1)
self.f2 = torch.nn.Linear(32 * h * w, 5)
def forward(self, x):
x = self.c1(x)
x = x.view(x.size(0), -1)
x = self.f2(x)
return x
I know the second derivative can be calculated by calling torch.autograd.grad()
twice, but the parameters in pytorch is organized by net.parameters()
, and I don't know how to compute the hessian for all parameters.
I have tried to use torch.autograd.functional.hessian()
in pytorch 1.5 as follows:
import torch
import numpy as np
from torch.nn import Module
import torch.nn.functional as F
class Net(Module):
def __init__(self, h, w):
super(Net, self).__init__()
self.c1 = torch.nn.Conv2d(1, 32, 3, 1, 1)
self.f2 = torch.nn.Linear(32 * h * w, 5)
def forward(self, x):
x = self.c1(x)
x = x.view(x.size(0), -1)
x = self.f2(x)
return x
def func_(a, b c, d):
p = [a, b, c, d]
x = torch.randn(size=[8, 1, 12, 12], dtype=torch.float32)
y = torch.randint(0, 5, [8])
x = F.conv2d(x, p[0], p[1], 1, 1)
x = x.view(x.size(0), -1)
x = F.linear(x, p[2], p[3])
loss = F.cross_entropy(x, y)
return loss
if __name__ == '__main__':
net = Net(12, 12)
h = torch.autograd.functional.hessian(func_, tuple([_ for _ in net.parameters()]))
print(type(h), len(h))
h
is a tuple, and the results are in strange shape. For example, the shape of \frac{\delta Loss^2}{\delta c1.weight^2}
is [32,1,3,3,32,1,3,3]
. It seems like I can combine them into a complete H
, but I don't know which part it is in the whole Hessian Matrix and the corresponding order.
Here is one solution, I think it's a little too complex but could be instructive.
Considering about these points:
torch.autograd.functional.hessian()
the first argument must be a function, and the second argument should be a tuple or list of tensors. That means we cannot directly pass a scalar loss to it. (I don't know why, because I think there is no large difference between a scalar loss or a function that returns a scalar)So here is the solution:
import torch
import numpy as np
from torch.nn import Module
import torch.nn.functional as F
class Net(Module):
def __init__(self, h, w):
super(Net, self).__init__()
self.c1 = torch.nn.Conv2d(1, 32, 3, 1, 1)
self.f2 = torch.nn.Linear(32 * h * w, 5)
def forward(self, x):
x = self.c1(x)
x = x.view(x.size(0), -1)
x = self.f2(x)
return x
def haha(a, b, c, d):
p = [a.view(32, 1, 3, 3), b, c.view(5, 32 * 12 * 12), d]
x = torch.randn(size=[8, 1, 12, 12], dtype=torch.float32)
y = torch.randint(0, 5, [8])
x = F.conv2d(x, p[0], p[1], 1, 1)
x = x.view(x.size(0), -1)
x = F.linear(x, p[2], p[3])
loss = F.cross_entropy(x, y)
return loss
if __name__ == '__main__':
net = Net(12, 12)
h = torch.autograd.functional.hessian(haha, tuple([_.view(-1) for _ in net.parameters()]))
# Then we just need to fix tensors in h into a big matrix
I build a new function haha
that works in the same way with the neural network Net
. Notice that arguments a, b, c, d
are all expanded into one-dimensional vectors, so that the shapes of tensors in h
are all two dimensional, in good order and easy to be combined into a large hessian matrix.
In my example, the shapes of tensors in h
is
# with relation to c1.weight and c1.weight, c1.bias, f2.weight, f2.bias
[288,288]
[288,32]
[288,23040]
[288,5]
# with relation to c2.bias and c1.weight, c1.bias, f2.weight, f2.bias
[32, 288]
[32, 32]
[32, 23040]
[32, 5]
...
So it is easy to see the meaning of the tensors and which part it is. All we need to do is to allocate a (288+32+23040+5)*(288+32+23040+5)
matrix and fix the tensors in h
into the corresponding locations.
I think the solution still could be improved, like we don't need to build a function works the same way with neural network, and transform the shape of parameters twice. But for now I don't have better ideas, if there is any better solution, please let me know.